Beyond the Buzzwords: What Actually Keeps Shots Coherent

A lot of AI filmmaking discussion focuses on the wrong part of this problem.

You hear about model routing, asset locking, production graphs, character consistency, and increasingly large multimodal models. Then there is the opposite argument: all of these products are just wrappers around somebody else's model, so the surrounding system does not really matter.

After building RanGen, I don't think either framing gets to the interesting part.

The question isn't how many models are involved.

If I ask for another shot in a scene, does it know which characters are there? What they're wearing? Where they're standing? What the room looks like? Which direction they're looking? What we established in the master?

And when I say, “Actually, put Mara in the black coat for this one,” does the system understand that as a deliberate creative change rather than continuity drift?

That is the problem we've been working on.

---

Filmmakers shouldn't have to prompt

There is an assumption buried in a lot of AI filmmaking tools that the filmmaker's job is to describe an image precisely enough for the model.

We don't think that is the right interaction.

When you're working with a cinematographer or storyboard artist, you don't normally say:

You say:

There is a huge difference between those two things.

The first is an instruction written for a generative model.

The second is direction.

For RanGen's Scene Generator, we want the filmmaker to work in the second mode. You talk to it the way you would talk to another filmmaker working with you. The system's job is to figure out the rest.

• → If Mara is already established, you should not need to redescribe her face.
• → If she's wearing the red jacket in this part of the script, you should not need to mention the jacket every time.
• → If the location has already been established, you shouldn't have to reconstruct the room in words every time you want another angle.

Those are production facts. The software should remember them.

---

The prompt still exists. It just shouldn't be the director's problem.

Obviously, generative models still need instructions.

Somewhere underneath the interface, something eventually has to tell an image or video model what to generate. But there is no reason that representation needs to be authored by the filmmaker.

This is an important distinction.

Those production instructions can include things the director never explicitly said because the project already knows them:

Then the system determines which of those things the generative model actually needs for this shot. That's a very different workflow from asking the user to become better at prompting.

---

A generated image is not production memory

Consider a simple scene.

Mara is standing in an alley wearing a red leather jacket.

The first image establishes dozens of things whether we consciously specified them or not: her appearance, the cut of the jacket, the shape of the alley, lighting direction, background details, camera height, perhaps even the weather.

Now I say:

I haven't mentioned the jacket. I haven't described the alley again. I haven't told RanGen what Mara looks like.

I shouldn't have to. The production already knows those things.

If the system relies only on my latest instruction, the model gets another opportunity to reinterpret everything.

• That's how the leather jacket becomes wool.
• That's how the alley suddenly gets a staircase.
• That's how a character's face gradually drifts over a sequence.

This isn't really a prompting problem. The information existed.

---

Drift and direction are not the same thing

Suppose Mara wears a red jacket in Scene 4 and a black coat in Scene 5.

There are two completely different reasons that might happen.

1. The model accidentally changed her clothes.
2. Or the director intentionally changed her clothes.

Pixels alone don't tell you which one happened. The production needs a record of intent.

In our case, that means keeping explicit state for things like characters, locations, costumes, props, scenes, and setups. The red jacket remains part of Mara's current look until something in the production changes it.

• Maybe the screenplay establishes the costume change.
• Maybe the system detects it automatically.
• Maybe the director simply says: "Actually, she's in the black coat now."

That is enough.

The filmmaker declares the creative change. RanGen updates the production state behind it. The filmmaker does not need to find a wardrobe node, edit a database relation, rewrite a prompt, or remember which reference images were used three shots ago.

That's our job.

---

What RanGen actually keeps track of

When a screenplay enters RanGen, we extract recurring production entities and relationships into the World Bible:

This gives us something much more useful than a long conversation history. We can answer boring but important questions:

• Who is actually in this shot?
• What is Mara wearing at this point in the story?
• Which environment does this take place in?
• Is there a prop that has to remain visible?
• Which setup are we shooting coverage from?
• What has already been established visually?

These sound like small details until you generate fifty shots. Then they become the difference between a sequence and fifty unrelated images.

---

More references can make the result worse

This was one of the things that became obvious while testing.

Suppose the master shot contains Mara, David, and Elena.

Now I ask:

The obvious thing would be to send the master plus references for Mara, David, and Elena into the next generation. But David and Elena aren't needed.

And depending on the image model, those extra identities can actually hurt the result. Features bleed. Wardrobe gets mixed. A visual detail belonging to one character appears on another.

So RanGen tracks who is physically present in the shot and filters the visual context accordingly.

If Mara is the only character in the close-up, the generator does not need everybody else's identity references competing for attention.

The filmmaker never has to manage that. They just asked for a close-up of Mara.

---

We ran into the same problem with wardrobe

Sending a model a character reference and a separate pile of clothing references sounds reasonable. Until there are multiple characters.

We know which outfit belongs to which person. The image model doesn't necessarily bind those inputs correctly every time.

So we can create composite Look Cards: a visual reference of a specific character in a specific active look.

Instead of asking the generator to infer:

we can give it a stronger condition:

It's a small architectural decision, but it removes one more thing the model has to guess. A lot of our work looks like that. Not trying to make the model smarter. Trying to reduce the number of opportunities it has to be wrong.

---

Every generation has a reference budget

You also cannot throw the entire project at a generative model every time.

Different models have different limits on visual inputs. And even when an endpoint technically accepts many references, that doesn't mean every reference contributes equally.

So RanGen has to decide what matters for the shot being requested.

If I say: "Give me a close-up of Mara holding the key."

For another shot, the priority could be completely different. That's why we think of references as a budget.

The Scene Generator interprets what the filmmaker is asking for, looks at what the production already knows, and assembles the most useful context it can for the underlying model.

Again, none of this should become work for the filmmaker.

---

Keeping the same face isn't enough

This is where AI “character consistency” demos often stop too early.

You can preserve Mara's face perfectly and still make unusable coverage. Imagine a simple dialogue scene.

The master establishes Sarah on the left side of the table and David on the right.

Now I say: "Give me Sarah's close-up."

Her face comes back perfectly. But she's looking camera-left. Or the window that was behind David is suddenly behind Sarah. Or David's shoulder appears on the wrong side of frame.

Each image can look excellent on its own. Cut them together and something feels wrong immediately.

That's because filmmaking continuity isn't just identity continuity. It's spatial.

The system needs some memory of blocking, screen direction, eyelines, environment, camera position, and what the setup already established.

---

This is why RanGen starts from setups and masters

We didn't invent this idea. Filmmakers have been using master shots forever.

The master establishes the space:

• Who is where?
• Which direction are they facing?
• Where is the camera?
• What does the environment look like?
• What's the basic lighting logic?

Then you shoot coverage relative to that.

RanGen's Scene Generator follows the same basic idea. The filmmaker can establish a setup and then simply direct coverage from it:

The system already has the scene context. Those instructions should behave like filmmaking instructions, not prompts that happen to produce pictures.

Underneath, RanGen can use the master, character state, location references, active looks, props, shot history, and other structured information to build the generation request. The director doesn't have to say any of that out loud.

Does it guarantee geometrically perfect coverage? No. Current generative models still struggle with exact spatial consistency.

But there is a big difference between asking a model to invent a scene from scratch for every shot and asking it to produce another shot from a scene the production has already established.

---

Previous shots help, but they aren't the truth

Another tempting approach is simply chaining generations.

For short sequences, this can work well. It also has a nasty failure mode.

A small mistake in B becomes conditioning for C. Then C introduces another small mistake. By D or E, the sequence can become very consistent with things nobody actually intended.

That's why we distinguish shot history from production state.

Previous approved shots are useful visual evidence. They are not the only source of truth.

Mara's identity should exist independently of whatever her face happened to look like in the most recent generation. The location should exist independently of a wall that accidentally moved in Shot 12.

The production remembers the world. The individual generations are attempts to render it.

---

We can't actually “lock” a probabilistic model

This is another place where the language around AI filmmaking gets misleading.

We can lock what gets sent to the model. We cannot completely control what the model does with them.

If Mara's correct Look Card, environment reference, prop, and setup are all supplied, the image model can still change the jacket collar. That's where the technology is right now.

So when we talk about continuity, we're not claiming deterministic pixels. We're talking about reducing unnecessary variance and making the intended production state explicit.

And when something is wrong, the filmmaker should be able to react like a filmmaker:

Not:

That difference in interaction matters to us as much as the underlying generation technology.

---

What we're actually trying to build

I don't know whether the future ends up being one giant filmmaking model or hundreds of specialized models behind the scenes.

I'm not sure it matters.

Models will change. The best image model six months from now probably won't be the one we use today.

And the filmmaker should be able to change any of those things by directing the system normally.

That's the layer we're building RanGen around.

Not prompting. Production memory plus direction.

The goal is for the technology underneath to become increasingly invisible, until working with the Scene Generator feels less like operating an AI tool and more like having another filmmaker beside you who already knows the project.